← All projects
early stage

Parti

An early-stage multi-agent orchestration system that runs a research, plan, implement, verify, and evaluate loop across a three-node home lab: a Mac Studio, a GB10 workstation, and a NAS.

The Parti workspace: sessions and agents on the left, the main agent’s thread in the middle with a plan running and a question for you, the project’s files on the rightThe Parti workspace: sessions and agents on the left, the main agent’s thread in the middle with a plan running and a question for you, the project’s files on the right

One task, start to finish

  1. 01You ask. The main agent plans the work and hands it to a coding worker: Claude Code or Codex.

    The plan for “Fix the export at 2x” running as step 3 of 5 on Codex, with a Claude Code worker and the files it readThe plan for “Fix the export at 2x” running as step 3 of 5 on Codex, with a Claude Code worker and the files it read

  2. 02A worker needs you: pick between two fixes, or approve a file write.

    Two questions waiting for you: which fix to take, and whether Claude Code may write to a fileTwo questions waiting for you: which fix to take, and whether Claude Code may write to a file

  3. 03An agent drives the built-in browser. You can watch it, take over or stop it.

    The main agent browsing a docs page about export scale, with its action log and a Stop buttonThe main agent browsing a docs page about export scale, with its action log and a Stop button

  4. 04Review the changes and the diff, then commit.

    The changes view: two changed files and the diff that stops applying the pixel ratio twice, ready to commitThe changes view: two changed files and the diff that stops applying the pixel ratio twice, ready to commit

  5. 05The team learns: it proposes new skills, which you accept, edit or reject.

    Proposals: a new skill, “Summarise a long thread”, with Accept, Edit first and RejectProposals: a new skill, “Summarise a long thread”, with Accept, Edit first and Reject

How it runs

Mac Studio · OrchestratorGB10 · Compute & HarnessResearchPlanImplementVerifyEvaluateNAS · Shared Workspace,Checkpoints & Provenance↺ feedback & iteration loop

Why build it

One person gets a coherent, capable team of AI collaborators without managing every agent or every conversation.

Agentic AIMulti-agent orchestrationModel evaluationContent provenanceLocal-first AI